Deep Interaction: An Efficient Human-AI Interaction Method for Large Reasoning Models
The paper proposes editing a model’s original chain of thought instead of asking it to regenerate from scratch.
Deep Interaction lets a user correct the specific faulty part of a reasoning trace while keeping the valid steps intact. The edited reasoning is distilled into a prompt that steers the model back onto the corrected path. On STEM reasoning tasks, the authors report more than a 25% gain in correction success rate and about 40% lower token use versus baseline interaction methods. ArXiv · AI/CL/LG's note
Deep Interaction lets a user correct the specific faulty part of a reasoning trace while keeping the valid steps intact. The edited reasoning is distilled into a prompt that steers the model back onto the corrected path. On STEM reasoning tasks, the authors report more than a 25% gain in correction success rate and about 40% lower token use versus baseline interaction methods. ArXiv · AI/CL/LG's note
score 4